Machine Learning Based Prediction of Anti-dengue Drugs Using Structure-activity Relationship Study

Dengue fever has emerged as a critical health threat in Bangladesh and recorded the highest mortality rate in 2023. It is also a global concern, impacting 390 million people annually. No FDA-approved drug is available yet to treat dengue. The Structure-activity relationship (SAR) can be a significant approach for predicting novel inhibitors and repurposing existing drugs against dengue. This study aims to develop a machine-learning model that can predict the pIC50 (M) value of novel inhibitors against the dengue NS2B-NS3 protein. In this connection, this methodology uses a structure-activity dataset of 1141 compounds. Each compound contains 1544 molecular descriptors (ID and 2D) considered as structural features and their effective percentage inhibition concentration at 50% (pIC50), against 'DENV-2 NS2B-NS3‘. This research implements 9 tree-based, and 9 other distinct machine learning algorithms considering all features to identify the best-performing algorithm. In the baseline test, the tree-based algorithms outperform the other conventional algorithms where the Extra Trees algorithm performs the best among the tree-based algorithms. Furthermore, the study utilizes the Random Forest algorithm to extract individual features' importance against pIC50. The features are fed into 6 different tree-based algorithms in the chronology of feature importance. The experiments depict an impressive result that some of the features have a negative impact on prediction. Eventually, the work eliminates 1502 molecular descriptors to differentiate only 42 important features that can produce the lowest RMSE of 0.3035, MSE of 0.0924 MAPE of 0.0515, and MAE of 0.2309. This optimized model has outperformed the baseline model evidently.